ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2016, Vol. 53 ›› Issue (11): 2567-2582.doi: 10.7544/issn1000-1239.2016.20150743

• 人工智能 • 上一篇    下一篇



  1. (浙江工业大学计算机科学与技术学院 杭州 310023) (
  • 出版日期: 2016-11-01
  • 基金资助: 
    国家自然科学基金项目(61602413,61379123,61502424);国家科技支撑计划基金项目(2012BAD10B01);浙江省自然科学基金项目(LY15F030014,LY15F020028) This work was supported by the National Natural Science Foundation of China (61602413, 61379123, 61502424), the National Key Technology R&D Program of China (2012BAD10B01), and the Natural Science Foundation of Zhejiang Province of China (LY15F030014, LY15F020028).

Kernel Sparse Representation Classification with Group Weighted Constraints

Zheng Jianwei, Yang Ping, Wang Wanliang, Bai Cong   

  1. (College of Computer Science and Technology, Zhejiang University of Technology, Hangzhou 310023)
  • Online: 2016-11-01

摘要: 提出了一种称为核加权组稀疏表示分类器(kernel weighted group sparse representation classifier, KWGSC)的新型模式分类算法. 通过在核特征空间而非原输入空间引入组稀疏性和保局性,KWGSC能够获得更有效的鉴别性重构系数用于分类表示. 为获得最优重构系数,提出了一种新的迭代更新策略进行模型求解并给出了相应的收敛性证明以及复杂度分析. 对比现存表示型分类算法,KWGSC具有的优势包括:1)通过隐含映射变换,巧妙地规避了经典线性表示算法所固有的规范化问题;2)通过联合引入距离加权约束和重构冗余约束,精确地推导出查询样本的目标类别标签;3)引入l\-2,p正则项调整协作机制中的稀疏性,获得更佳的分类性能. 人造数值实验表明:经典线性表示型算法在非范数归一化条件下无法找到正确的重构样本,而KWGSC却未受影响. 实际的公共数据库验证了所提分类算法具有鲁棒的鉴别力,其综合性能明显优于现存算法.

关键词: 稀疏表示技术, 保局性, 组稀疏正则项, 核技术, 范数归一化问题

Abstract: A new classification method called KWGSC (kernel weighted group sparse representation classifier) is proposed for pattern recognition. KWGSC integrates both group sparsity and data locality in the kernel feature space rather than in the original feature space. KWGSC can learn more discriminating sparse representation coefficients for classification. The iteratively update solution of the l\-2,p-norm minimization problem for KWGSC is also presented. There are several appealing aspects associated with KWGSC. Firstly, by mapping the data into the kernel feature space, the so-called norm normalization problem that may be encountered when directly applying sparse representation to non-normalized data classification tasks will be naturally alleviated. Secondly, the label of a query sample can be inferred more precisely by using of distance constraints and reconstruction constraints in together. Thirdly, the l\-2,p regularization (where p∈(0,1]) is introduced to adjust the sparsity of collaborative mechanism for better performance. Numeric example shows that KWGSC is able to perfectly classify data with different normalization strategy, while conventional linear representation algorithms fail completely. Comprehensive experiments on widely used public databases also show that KWGSC is a robust discriminative classifier with excellent performance, being outperforming other state-of-the-art approaches.

Key words: sparse representation (SR), locality-constraint, group sparse regularizer, kernel trick, norm normalization problem